Masked Face Detection Based On Locally Nonlinear Feature Fusion

Masked face detection is one of the most important problems in face de-tection currently.In the era of artificial intelligence,face detection has been widely used as an important research direction in the field of image process-ing and computer vision.However,the requirements for specific and com-plex practical application scenarios are also increasing.This paper makes an in-depth study on the face detection algorithm in color images,and proposes a masked face detection algorithm based on locally nonlinear feature fusion.LNFF-Net(Locally Nonlinear Feature Fusion-based Network)is divided into two stages:candidate region generation and candidate region discrimination,which mainly solve the following problems.(1)This algorithm uses the pre-trained model to extract network learn-ing features for the problem of insufficient training data.(2)In view of the problem of missed detection and misdetection of masked face candidate regions,the GBVS model is introduced to calculate visual saliency maps,and it is proposed to use traditional underlying visual features to modify the classification score on the network feature map.(3)Aiming at the problem of single feature fusion strategy,a new non-linear fusion method is proposed to highlight the feature from the face region and suppress the background region.(4)In response to the problem of repeatedly extracting features from the traditional candidate region discriminant structure,the Fast R-CNN model is introduced and adjusted to extract all regional features at once,reducing the time complexity and improving the robustness of masked face features.In this paper,the MAFA and COFW public masked face datasets in real application scenarios are used in experiments.The results prove that the nonlinear fusion method in the LNFF-Net algorithm has a higher recall rate and F1 value than other fusion methods.After feature fusion,the gener-ated candidate regions are positioned more accurately and more compactly.Compared with other detection algorithms,the average accuracy of detecting masked faces,especially severely occluded faces is significantly improved.